Crowd Counting Based on Multiresolution Density Map and Parallel Dilated Convolution

Abstract

The current crowd counting tasks rely on a fully convolutional network to generate a density map that can achieve good performance. However, due to the crowd occlusion and perspective distortion in the image, the directly generated density map usually neglects the scale information and spatial contact information. To solve it, we proposed MDPDNet (Multiresolution Density maps and Parallel Dilated convolutions’ Network) to reduce the influence of occlusion and distortion on crowd estimation. This network is composed of two modules: (1) the parallel dilated convolution module (PDM) that combines three dilated convolutions in parallel to obtain the deep features on the larger receptive field with fewer parameters while reducing the loss of multiscale information; (2) the multiresolution density map module (MDM) that contains three-branch networks for extracting spatial contact information on three different low-resolution density maps as the feature input of the final crowd density map. Experiments show that MDPDNet achieved excellent results on three mainstream datasets (ShanghaiTech, UCF_CC_50, and UCF-QNRF).